Skip to content

Setup CNTK on Linux

Jump to bottom Edit New page
Mark Hillebrand edited this page Jun 10, 2016 · 43 revisions

CNTK Production Build and Test configuration

CNTK may be successfully run in many Linux configurations, but in case you want to avoid possible compatibility issues you may get yourself familiar with CNTK Production Build and Test configuration where we list all dependency component and component versions that we use.

CNTK as a Docker container

Before moving any further you may consider deploying CNTK as Docker container. Read the corresponding section.

Current limitations and precautions

Please, read carefully this section before you proceed with your system configuration. The information below may save you a lot of time otherwise spent on build errors debugging.

Expected component locations in configure and Makefile scripts

Today configure and Makefile scripts support only limited set of installation paths for all dependency components listed in this section. We know, that this is a limitation and will fix it soon (also if you feel like improving these scripts yourselves and submit your proposed changes your help is welcome and much appreciated).

configure looks for all dependency components among the paths listed in default_path_list variable defined within the script.

IMPORTANT! If you want to modify default_path_list variable in configure to add a custom path for a certain dependancy component be sure to check the correspondent section of Makefile. Otherwise you may get build errors due to inability of finding INCLUDE files, libraries, etc.

Installation methods and paths of dependency components

Below we list all dependency components required to build CNTK and explain how to install them. We understand that there are many other ways to get the same components. However, if you prefer an alternative way of installation, please ensure that you get the same thing, because quite often alternative installation sources, namely network distibution packagespackages (like Debian, RPM, etc.) contain older versions of the software, miss some libraries, etc. In some sections below we specifically highlight these limitations, but please take it as a general precaution.

Using of make -j

In most of the sections we suggest using make -j command to invoke parallel build jobs and thus increasing the speed of the build process. However please be aware that on some systems and especially on virtual machines using make -j may result in "Out of memory" errors. If you face this, just use "plain" make or limit the number of jobs that run simultaneously (two simultaneous jobs usually work for the most of the systems - use the command make -j 2).

Simultaneous installation of different versions of the same development packages

Be very careful in case you would like to have several installations of some of the development packages mentioned below on the same system. It may result in very hard to debug build errors as you can see in this post.

And now let's proceed to the setup.

If you would like to know what prerequisite configuration is used in the CNTK production environment, i.e. what we use internally for building and testing, see this section

64-bit OS

You need a 64-bit Linux installation to use CNTK.

C++ Compiler

Ensure your installation has a C++ compiler. Many distributions do not include it by default. Refer to your platform documentation on how to check for and obtain a C++ compiler.

Example: for Ubuntu, run the following command:

dpkg --list | grep compiler

if in the output you do not see something like g++-4.8 4.8.4-2ubuntu1~14.04 amd64 GNU C++ compiler then a C++ compiler is not installed. Install it with:

sudo apt-get install g++

Git

Install Git on your system as described here.

ACML

Install the AMD Core Math Library (ACML):

tar -xzvf ./acml-5-3-1-ifort-64bit.tgz
sudo ./install-acml-5-3-1-ifort-64bit.sh
  • Accept the default installation path
  • Add the following environment variable to your current session and your .bashrc profile:
export ACML_FMA=0
export LD_LIBRARY_PATH=/opt/acml5.3.1/ifort64/lib:/opt/acml5.3.1/ifort64_mp/lib:$LD_LIBRARY_PATH

OPTIONAL. MKL

An alternative to the ACML library is the Intel Math Kernel Library (Intel MKL). CNTK supports using the Intel MKL via a custom library version ("CNTK custom MKL").

IMPORTANT You can NOT directly build CNTK using a regular installation of Intel MKL SDK; you can however, build your own CNTK custom MKL to link against, see here.

  • Create a directory on your machine to hold CNTK custom MKL versions, e.g.:
sudo mkdir /usr/local/CNTKCustomMKL
  • Download the required CNTK custom MKL from the CNTK web site. Un-tar it into your CNTK MKL path, creating a numbered sub directory within. For example, if you are on latest master, download CNTKCustomMKL-Linux-1.tgz and extract its contents to /usr/local/CNTKCustomMKL/1:
wget https://www.cntk.ai/mkl/CNTKCustomMKL-Linux-1.tgz
sudo tar -xzf CNTKCustomMKL-Linux-1.tar.gz -C /usr/local/CNTKCustomMKL

Note: if you want to build other CNTK source versions, you may need to install alternate CNTK custom MKL versions. Check the variable cntk_custom_mkl_version in CNTK's configuration script configure to determine which.

  • When configuring the build (cf. below), specify the option --with-mkl or --with-mkl=<directory>, e.g., --with-mkl=/usr/local/CNTKCustomMKL.

Open MPI

Install Open Message Passing Interface Library (Open MPI).

We recommend installing from sources as described below because a lot of distribution packages contain older versions and miss the libraries required by CNTK. Current CNTK Open MPI version requirement is at least 1.10. Please, check whether you have older version installations on your system and if you do, either deinstall them or ensure (via, e.g. symbolic links) that CNTK build procedure is using the required version. Otherwise you may get hard to debug build errors as you can see in this post.

  • Get the installation sources:
wget https://www.open-mpi.org/software/ompi/v1.10/downloads/openmpi-1.10.1.tar.gz
  • Unpack, build and install Open MPI (to /usr/local/mpi in this example):
tar -xzvf ./openmpi-1.10.2.tar.gz
cd openmpi-1.10.2
./configure --prefix=/usr/local/mpi
make -j all
sudo make install
  • Add the following environment variable to your current session and your .bashrc profile (Prepending the new path, ensures this version is used as opposed to a default version available through the OS):
export PATH=/usr/local/mpi/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/mpi/lib:$LD_LIBRARY_PATH

GPU Specific Packages

This section outlines the packages you need to setup in order for CNTK to leverage GPUs.

If you do NOT intend to use any GPUs with CNTK and will rely on the CPU only, you may skip directly to the next general section. Otherwise proceed further.

Checking your GPU compatibility and getting the latest driver

You need a CUDA-compatible graphic card available to use CNTK GPU capabilities. You can check whether your card is CUDA-compatible here and here (for older cards).

If you have the required card, install the latest driver:

  • Select your card and download the driver pack from here (usually available as .run file)
  • If your X Windows manager is running, the driver installation will likely fail. Open a remote terminal session to your machine and stop the X Windows manager. Refer to your platform documentation for the exact commands.

Example: for Ubuntu, use the following command (in case of lightdm as your X Windows manager):

sudo stop lightdm
  • Install the driver as in the example below (note that the file name may be different for your system):
sudo chmod +x ./NVIDIA-Linux-x86_64-352.63.run
sudo ./NVIDIA-Linux-x86_64-352.63.run

We recommend accepting the default installation options

Note that the driver installation program may complain about nouveau kernel driver. Refer to your platform documentation for instructions for disabling it. For Ubuntu you may use this set of instructions.

  • If stopped during the steps above, start X Windows manager. Refer to your platform documentation for the exact commands.

Example: for Ubuntu, use the following command (in case of lightdm as your X Windows manager):

sudo start lightdm

CUDA 7.5

IMPORTANT. Currently CNTK is set to be built with CUDA 7.5. Building it with higher versions of CUDA may result in build errors.

The procedure below explains how to install CUDA using .run file distribution. You can also use DEB or RPM packages installation. You will find the package for your system at NVIDIA CUDA 7.5 Download page and installation instructions in CUDA Online Documentation.

Download and install the NVIDIA CUDA 7.5 Toolkit:

  • Find the .run file for your platform here and download it.

  • If your X Windows manager is running, the installation will likely fail. Open a remote terminal session to your machine and stop the X Windows manager. Refer to your platform documentation for the exact commands.

Example: for Ubuntu use the following command (in case of lightdm as your X Windows manager):

sudo stop lightdm
  • Install CUDA 7.5 Toolkit (note that the .run file name may be different for your system):
sudo chmod +x ./cuda_7.5.18_linux.run
sudo ./cuda_7.5.18_linux.run

IMPORTANT! When prompted by the installer:

Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 352.39? ((y)es/(n)o/(q)uit):

select no if you have already installed the latest driver by performing the steps in the previous section. If you have not done it, select yes, but we strongly recommend updating to the latest driver after installing CUDA toolkit.

If you declined the driver installation from the CUDA 7.5 package as described in the previous section, you will get the following warning at the end of the installation:

***WARNING: Incomplete installation! This installation did not install the CUDA Driver. A driver of version at least 352.00 is required for CUDA 7.5 functionality to work.
To install the driver using this installer, run the following command, replacing <CudaInstaller> with the name of this run file:
    sudo <CudaInstaller>.run -silent -driver

Ignore it

  • If stopped during the steps above, start X Windows manager. Refer to your platform documentation for exact commands.

Example: for Ubuntu use the following command (in case of lightdm as your X Windows manager):

sudo start lightdm
  • Add the following environment variable to your current session and your .bashrc profile (if you modified the default paths during the CUDA installation, change the values below accordingly):
export PATH=/usr/local/cuda-7.5/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64:$LD_LIBRARY_PATH

The next step is optional. You may skip further to the next section.

OPTIONAL. Verifying CUDA 7.5 installation

You may verify your CUDA installation by compiling the CUDA samples (we assume the default paths were used during the CUDA installation). Note that building all samples is a lengthy operation:

cd ~/NVIDIA_CUDA-7.5_Samples/
make

After the successful build invoke deviceQuery utility:

~/NVIDIA_CUDA-7.5_Samples/1_Utilities/deviceQuery/deviceQuery

If everything works well, you should get an output similar to the one below:

/home/alexey/NVIDIA_CUDA-7.5_Samples/1_Utilities/deviceQuery/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "Quadro 600"
  CUDA Driver Version / Runtime Version          8.0 / 7.5
  CUDA Capability Major/Minor version number:    2.1
  Total amount of global memory:                 1016 MBytes (1065734144 bytes)
  ( 2) Multiprocessors, ( 48) CUDA Cores/MP:     96 CUDA Cores
  GPU Max Clock rate:                            1280 MHz (1.28 GHz)
  Memory Clock rate:                             800 Mhz
  Memory Bus Width:                              128-bit
  L2 Cache Size:                                 131072 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(65536), 2D=(65536, 65535), 3D=(2048, 2048, 2048)
  Maximum Layered 1D Texture Size, (num) layers  1D=(16384), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(16384, 16384), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total number of registers available per block: 32768
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  1536
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (65535, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 1 copy engine(s)
  Run time limit on kernels:                     No
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  Device supports Unified Addressing (UVA):      Yes
  Device PCI Domain ID / Bus ID / location ID:   0 / 2 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 8.0, CUDA Runtime Version = 7.5, NumDevs = 1, Device0 = Quadro 600
Result = PASS

GPU Deployment kit

If you installed CUDA using DEB or RPM package (see previous section), you already have GPU Deployment kit installed, so skip to the next section. Otherwise proceed further.

Get and install the latest version of NVIDIA GPU Deployment kit:

  • Find the .run file for your platform here. Select the latest version. Download the file either with the browser or copy the link to the required file and use ```wget`` command as shown in the example below (note that the file name may be different for your system):
wget http://developer.download.nvidia.com/compute/cuda/7.5/Prod/gdk/gdk_linux_amd64_352_79_release.run
  • If your X Windows manager is running, the installation will likely fail. Open a remote terminal session to your machine and stop the X Windows manager. Refer to your platform documentation for the exact commands.

Example: for Ubuntu use the following command (in case of lightdm as your X Windows manager):

sudo stop lightdm
  • Install the GPU Deployment kit (note that the .run file name may be different for your system):
sudo chmod +x ./gdk_linux_amd64_352_79_release.run
sudo ./gdk_linux_amd64_352_79_release.run

We recommend accepting the default values during installation.

  • If stopped during the steps above, start X Windows manager. Refer to your platform documentation for exact commands.

Example: for Ubuntu use the following command (in case of lightdm as your X Windows manager):

sudo start lightdm

CUB

Get and install NVIDIA CUB.

IMPORTANT! Install NVIDIA CUB using the exact version and target path as specified below. This is necessary because it is expected by the CNTK build configuration program.

Use the following commands:

wget https://github.com/NVlabs/cub/archive/1.4.1.zip
unzip ./1.4.1.zip
sudo cp -r cub-1.4.1 /usr/local

cuDNN

If you would like to use CNTK together with the NVIDIA CUDA Deep Neural Network library (cuDNN), you'll need to get and install the cuDNN library during this step.

IMPORTANT! Install cuDNN using the exact version and target path as specified below. This is necessary because it is expected by the CNTK build configuration program.

  • Use the following commands:
wget http://developer.download.nvidia.com/compute/redist/cudnn/v4/cudnn-7.0-linux-x64-v4.0-prod.tgz
tar -xzvf ./cudnn-7.0-linux-x64-v4.0-prod.tgz
sudo mkdir /usr/local/cudnn-4.0
sudo cp -r cuda /usr/local/cudnn-4.0
  • Add the following environment variable to your current session and your .bashrc profile:
export LD_LIBRARY_PATH=/usr/local/cudnn-4.0/cuda/lib64:$LD_LIBRARY_PATH

This completes GPU Specific section

OPTIONAL. OpenCV

If you want to build the CNTK Image Reader, you need to install Open Source Computer Vision (OpenCV).

Some aspects to consider:

  • OpenCV requires at least 5.5 GB of free disk space
  • Building OpenCV is a lengthy process and can take a couple of hours or even more
  • If you plan to use both CUDA and OpenCV, install CUDA first

OpenCV can have many interfaces and options. In this section we cover only parts necessary to build CNTK. Read more about installing OpenCV here and more generally here.

IMPORTANT! Install OpenCV using the exact version and target path as specified below. This is necessary because it is expected by the CNTK build configuration program.

  • Install OpenCV prerequisites. See OpenCV installation instructions for the full package list and explanations. See your platform documentation on how to install the packages.

Example: for Ubuntu use the following command:

sudo apt-get install cmake libgtk2.0-dev pkg-config libavcodec-dev libavformat-dev libswscale-dev
  • Get OpenCV and install it. Use the following commands:
wget https://github.com/Itseez/opencv/archive/3.0.0.zip
unzip 3.0.0.zip
cd opencv-3.0.0
mkdir release
cd release
cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local/opencv-3.0.0 ..
make all
sudo make install

Note that in the instructions above we suggest using "plain" make rather than make -j. We found that using make -j with OpenCV results in unstable system behavior and may result in a build failure and system crash.

OPTIONAL. zlib and libzip

If you want to build the CNTK Image Reader, that supports ZIP File format, you need to install zlib and libzip libraries.

Please note, that you also need to install OpenCV to build CNTK Image Reader. See previous section.

Getting zlib

zlib today is a part of many popular Linux distributions so with the high probability you have it preinstalled. However note, that libzip requires zlib version 1.1.2 or later and this usually is NOT preinstalled.

See your platform documentation on how to get the required zlib package or get it directly from zlib website.

Example: for Ubuntu use the following command:

sudo apt-get install zlib1g-dev

Building libzib

libzip is available in different network distribution packages, but we found that many of them contain old versions. Using these versions will likely result in build errors. So we strongly recommend to build libzip from sources as described below.

Note that the following procedure will install libzip to /usr/local. This is exactly where CNTK build procedure expects it (see the beginning of this page for details). If you want to install libzip to a different path see instructions in INSTALL file in the root of libzip distribution folder. However beware that in this case you have to manually edit configure AND Makefile of CNTK to support this path.

Use the following commands:

wget http://nih.at/libzip/libzip-1.1.2.tar.gz
tar -xzvf ./libzip-1.1.2.tar.gz
cd libzip-1.1.2
./configure
make -j all
sudo make install

Add the following environment variable to your current session and your .bashrc profile:

export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

Getting CNTK Source code

Before proceeding further, please note, that if you plan on making modifications to the CNTK code you should read the information on Developing and Testing in this Wiki.

Use Git to clone the CNTK Repository and access the source code:

git clone https://github.com/Microsoft/cntk

IMPORTANT! The command above does not provide you with the access to 1-bit Stochastic Gradient Descent (1bit-SGD) component. If you want to build CNTK with 1bit-SGD functionality enabled, please read the instructions on this page and then proceed with installation.

Building CNTK

To build CNTK use the following commands (we assume that the CNTK repository was cloned to ~/Repos/cntk):

cd ~/Repos/cntk
mkdir build/release -p
cd build/release
../../configure

Ensure that the configure output corresponds to the packages you installed in the previous sections. I.e. ensure that configure finds CUDA if installed, etc.

make -j all

This should produce a release build of CNTK. In case you would like to get a debug build use the following parameter when invoking configure:

../../configure --with-buildtype=debug

Quick test of CNTK build functionality

To ensure that CNTK is working properly in your system, you can quickly run Simple2d example from CNTK Examples set. This example trains a simple network and can be directed to use either CPU or GPU, which helps quickly ensure that CNTK is functioning properly.

Below we assume that the CNTK repository is cloned to ~/Repos/cntk and build/release was used as a sub-directory for the build.

  • Provide the path to the CNTK binaries and switch to the Simple2d directory:
export PATH=$HOME/Repos/cntk/build/release/bin:$PATH
cd ~/Repos/cntk/Examples/Other/Simple2d/Data

First try the example:

cntk configFile=../Config/Simple.cntk

If the sample runs, i.e. if there are no error messages, you will get output related first to reading the configuration, followed by the output of the actual network training. Running the command a second time will report that the model has already been trained. To perform the training again, you must first remove the existing models by using the following command:

rm ../Output/Models/*

Trying CNTK with CPU

As seen previously, the Simple2d example produces an extensive output. Among this output is information regarding what processor was used for the Network Builder. If you perform the following set of commands:

cntk configFile=../Config/Simple.cntk &> out
cat out | grep Builder

you should get the following output

    SimpleNetworkBuilder = [
    SimpleNetworkBuilder = [
SimpleNetworkBuilder Using CPU

The last string is the informative one for our purposes. It tells us that CNTK was indeed using CPU.

Trying CNTK with GPU

If you build CNTK for GPU usage, try using GPU by executing the following commands:

cntk configFile=../Config/Simple.cntk deviceId=auto &> out
cat out | grep Builder

Expected output in this case is:

    SimpleNetworkBuilder = [
    SimpleNetworkBuilder = [
SimpleNetworkBuilder Using GPU 0

Note that GPU id may be different. deviceId parameter defines what processor to use for Network Builder:

  • deviceId=-1 means use CPU. Default value
  • deviceId=X where X is an integer >=0 means use GPU X, i.e. deviceId=0 means GPU 0, etc.
  • deviceId=auto means use GPU, select GPU automatically

Next steps

Add a custom footer

Getting started

Configuring CNTK

Describing Networks with BrainScript

Data readers

Evaluating CNTK Models

Advanced topics

Licenses

Clone this wiki locally